Federated learning, ethics, and the double black box problem in medical AI